Issue 6 — 12/05/2026

Perplexity crosses 45 M monthly active users and 1 B queries/month in 2026

Second Talent · 09/05/2026

Perplexity AI now serves more than 45 million monthly active users — more than double the 22 million recorded in early 2025 — and processes over one billion queries per month, according to DemandSage figures cited by Second Talent. The platform responds to 99.95% of queries versus Google AI Overviews' approximately 55% coverage (SE Ranking, 2025). For GEO practitioners, Perplexity is no longer a secondary target: at this query volume, citation slots inside Perplexity answers represent a material acquisition channel, particularly for B2B and technical audiences.

Perplexity's 'Computer' orchestrates 19 AI models via dynamic sub-agent architecture

Zen van Riel — AI Engineer Blog · 25/02/2026

Launched 25 February 2026, Perplexity Computer routes tasks across 19 AI models — with Claude Opus 4.6 as the core reasoning layer — through dynamically spawned sub-agents, and connects to 400-plus app integrations at $200/month (Perplexity Max tier). The orchestration-layer design means retrieval, reasoning, and synthesis are handled by specialised models rather than a single system, increasing the premium placed on structured, entity-rich source content that survives multi-hop retrieval. Practitioners should ensure their content is schema-annotated and entity-consistent so it is legible to retrieval components across heterogeneous model pipelines.

Perplexity error rate on PPC queries half that of Google AI Overviews — WordStream 2026

TechShali · 12/05/2026

An independent WordStream (2026) benchmark cited by TechShali found Perplexity returned a 13% error rate on pay-per-click questions versus a 26% error rate for Google AI Overviews on the same query set. AI search traffic grew 527% year-over-year between January–May 2024 and January–May 2025, per Semrush data also cited in the article. Accuracy differentials of this magnitude influence which platform professionals trust for research, reinforcing the case for allocating GEO effort across both Perplexity and Google AI Overviews rather than treating them as equivalent channels.

LLM pre-training pipeline: raw web text to aligned model — practitioner walkthrough

Medium / Ali (salisai) · 01/05/2026

A detailed first-person account of building an LLM from scratch describes the full data pipeline: mass web-text collection, deduplication, quality filtering, tokenisation, pre-training, instruction fine-tuning, and RLHF alignment. The post explains how decisions made at the data-collection stage propagate through every subsequent model behaviour. For GEO practitioners, this is a useful reminder that content excluded by quality filters or robots.txt at crawl time never reaches the model's parametric memory — making crawlability and demonstrated factual rigour prerequisites for organic LLM citation, not optional enhancements.

AI referrals to top websites spiked 357% YoY in June 2025, reaching 1.13 billion visits

AuraSearch · 12/05/2026

AuraSearch reports AI referral traffic to top websites grew 357% year-over-year in June 2025, reaching 1.13 billion visits, while nearly 29% of buyers now use AI-powered search tools more frequently than traditional Google queries. The same source states that content containing verifiable data earns 30%–40% more visibility in LLM-generated answers than purely qualitative content, and that approximately 60% of searches end without a click due to Google AI Overviews. Practitioners should prioritise embedding quantified, source-attributed claims in content rather than qualitative assertions, as data density appears to be a measurable citation signal.

67% of Google AI Overview citations reward five specific content formats — Wellows 2026

Wellows · 03/05/2026

Wellows reports (note: internal dataset, methodology not independently verified) that 67% of Google AI Overview citations favour five specific content formats, yet only 2.1% of audited pages deploy all five — indicating a structural supply gap in citation-ready content. The finding, if directionally accurate, suggests that format compliance alone may offer competitive differentiation even without domain authority advantages. Practitioners should treat this as a prompt to audit existing content against structured-format criteria (FAQ schema, HowTo markup, modular answer blocks) rather than a confirmed benchmark. (Unconfirmed — single-source internal data)

Benchmarking Retrieval Strategies for Biomedical Retrieval-Augmented Generation: A Controlled Empirical Study

Bal, D.P. and Puhan, S. · https://arxiv.org/abs/2605.02520v1 · 04/05/2026

Using a fixed generation model (GPT-4o-mini), shared vector store (ChromaDB), and identical embeddings, the authors isolate the effect of five retrieval strategies — Dense Vector Search, Hybrid BM25+Dense, Cross-Encoder Reranking, Multi-Query Expansion, and Maximal Marginal Relevance — on RAG answer quality in a biomedical QA pipeline. Because the generation layer is held constant, observed performance differences are attributable solely to retrieval design. (Pre-publication / arXiv) For GEO practitioners, the practical implication is structural: AI platforms that implement hybrid or reranking retrieval will surface content that is both semantically relevant and keyword-precise, meaning pages optimised for entity clarity and exact-match terminology alongside semantic depth are best positioned across retrieval architectures.

A Hybrid Retrieval and Reranking Framework for Evidence-Grounded Retrieval-Augmented Generation

Irany, F.A. and Akwafuo, S. · https://scirate.com/arxiv/2605.01664 · 01/05/2026

This paper presents a citation-aware RAG framework using Amazon Bedrock Knowledge Bases for ingestion, chunking, embedding, and retrieval, with a reranking layer that scores passages by evidence relevance before generation — specifically targeting biomedical and healthcare document QA. The framework's explicit citation-verification step means generated claims must be traceable to retrieved source passages. (Pre-publication / arXiv) For GEO practitioners, the design pattern is instructive: as AI answer systems increasingly incorporate citation-grounding and evidence reranking, content structured as discrete, self-contained, claim-per-passage blocks — rather than flowing prose — is mechanically better suited to being retrieved, reranked highly, and cited.